On April 13, Chief Judge Richard Seeborg of the U.S. District Court for the Northern District of California refused to get rid of a case OpenAI badly wanted gone. The order runs five pages and turns almost entirely on an obscure jurisdictional doctrine. It is also the clearest signal yet that the chatbot-harm docket will not be disposed of on procedure.
The case is Lyons v. OpenAI Foundation, No. 3:25-cv-11037, filed in December 2025 by Emily Lyons, administrator of the estate of Stein-Erik Soelberg. Soelberg killed his 83-year-old mother, Suzanne Adams, at their home in Old Greenwich, Connecticut, and then himself, after months of conversations with ChatGPT. The complaint names OpenAI, its affiliates, CEO Sam Altman, and certain employees and investors; it alleges GPT-4o reinforced Soelberg's paranoid delusions rather than interrupting them, and that OpenAI "rushed through or ignored" internal safety protocols before shipping it.
OpenAI moved in March to dismiss or stay under the Colorado River abstention doctrine, arguing a parallel state case — filed by First County Bank, executor of Adams' estate — covered the same ground. "Even though the cases share certain key facts and issues, there is doubt that resolution of the state court proceedings will resolve this matter, and so a Colorado River stay or dismissal is not warranted," Seeborg wrote. He also found OpenAI had "repeatedly misstat[ed] the law" on whether the doctrine compels abstention. It does not; it permits it.
The claims that survived, and why the reasoning matters
Both estates plead the same six counts: strict liability and negligence for design defect and failure to warn, a California Unfair Competition Law claim, wrongful death, and a survival action. That is a product-liability complaint, not a content complaint — a distinction doing enormous work across this docket. Seeborg's reasoning matters more than his outcome. He drew a line between two theories of defect in the same model. "Similarly, failure to warn depends on defendants' knowledge, actual or constructive, of the risk which may not be the same risk in the case of Soelberg's violence towards himself as compared to towards Adams," he wrote. Whether a chatbot is defectively designed as to third-party harm is a separate factual question from whether it is defectively designed as to self-harm, and a duty to warn may attach differently to each. That is standard products reasoning applied to a language model — and it implies discovery into what OpenAI knew, when, and about which failure mode.
That discovery is now happening at scale. In February, San Francisco Superior Court coordinated twelve actions as In re: ChatGPT Product Liability Cases, JCCP No. 5431, before Judge Ethan P. Schulman in Department 304. Schulman has since approved a plaintiffs' leadership slate of four co-leads, a liaison counsel, and a six-member steering committee — the scaffolding of mass tort litigation, not one-off suits. Raine v. OpenAI, No. CGC-25-628528, brought in August 2025 by the parents of 16-year-old Adam Raine, sits inside it. An October 2025 amended complaint added intentional-misconduct allegations and a punitive damages demand; OpenAI answered that November, denying liability and arguing the harm flowed from misuse in violation of its terms of use. Its lawyers wrote that "Adam Raine's death is a tragedy" while contesting causation. No trial date is set.
The docket keeps growing, and not only in private hands. On June 1, 2026, Florida became the first state to sue OpenAI and Altman directly, alleging in an 83-page complaint an "insatiable quest to win the AI arms race and amass large fortunes, despite knowing the danger of ChatGPT." The Social Media Victims Law Center and the Tech Justice Law Project, which filed seven of the coordinated suits in November 2025, keep filing more.
"These lawsuits are about accountability for a product that was designed to blur the line between tool and companion all in the name of increasing user engagement and market share," said Matthew P. Bergman, SMVLC's founding attorney.
Why Section 230 is a weak shield here
The industry's instinct is to reach for Section 230, which immunizes services from liability for content provided by another information content provider. The problem is structural: when a model generates the output, there is no third party, and providers have always remained liable for content they create in whole or in part. Plaintiffs have also pleaded around the statute by suing over design — engagement optimization, sycophancy, degraded crisis routing, inadequate warnings — rather than over any particular string of text.
The First Amendment defense has been tested and lost once already. In May 2025, U.S. District Judge Anne Conway declined to hold Character.AI's outputs protected speech at the motion-to-dismiss stage in Garcia v. Character Technologies, No. 6:24-cv-01903 (M.D. Fla.), letting most product-liability and consumer-protection claims proceed. Character.AI and Google settled that case and four related ones in January 2026 — avoiding appellate review of the holding.
The industry-wide implication is straightforward and expensive. If a model is a product, design choices face risk-utility balancing, safer alternative designs become a jury question, internal red-team findings become exhibits, and failure to warn attaches to known risks shipped anyway. Insurers price that differently than content moderation.
What to watch
Three things. First, the demurrers in JCCP 5431 — the coordination order contemplates OpenAI raising identical threshold defenses across all cases, and Schulman's ruling will be the first real appellate-track test of Section 230 and First Amendment defenses to generated output. Second, whether federal cases like Lyons proliferate enough to draw a petition to the Judicial Panel on Multidistrict Litigation; there is no MDL yet, and the split between state-court coordination and scattered federal filings is unstable. Third, trial settings: none of the marquee cases has one, and the first firm date resets settlement math docket-wide.
Be careful with the volume figures circulating. Bloomberg Businessweek has tracked roughly 40 suits against chatbot makers since late 2024; plaintiffs' firms say they are investigating dozens more. Those are intake numbers, not filings. No court, tracker, or named attorney has published a forecast that the count will double to a specific figure — it should not be reported as one.
---
This story covers suicide and death. If you are struggling, free and confidential support is available. In the U.S., call or text 988 to reach the Suicide and Crisis Lifeline. Outside the U.S., you can find a local helpline at findahelpline.com.
“Even though the cases share certain key facts and issues, there is doubt that resolution of the state court proceedings will resolve this matter, and so a Colorado River stay or dismissal is not warranted.”— Richard Seeborg, Chief Judge, U.S. District Court, N.D. Cal.